Dwarkesh Podcast AI researchers debate how close we are to recursive self-improvement
2556 snips
Sep 11, 2026 
Charlie O’Neill, Base10’s model-training lead; John Schulman, OpenAI co-founder and Thinking Machines chief scientist; and Beren Millidge, Zyphra’s CTO, debate how close AI is to recursive self-improvement. They explore why progress might stall, the challenges of continual learning, simulation-to-reality gaps, reinforcement learning, deployment data, and whether autonomous AI workers could soon tackle months-long knowledge work.
AI Snips
Chapters
Transcript
Episode notes
How AI Research Agents Will Be Trained
- AI researchers will likely train research agents through human feedback plus diverse multi-step practice environments.
- Labs may preserve frontier knowledge by turning recent bugs and discoveries into environments, continually distilling the latest research back into models.
Why Labs Keep Adding Domain-Specific Skills
- Labs are scaling long-horizon RL across coding, finance, presentations, and eventually the wider economy to cultivate persistence and task management.
- Domain-specific training remains useful because it improves runtime efficiency even if models could learn domains in context.
Sim-To-Real Depends On Sample Efficiency
- Simulated environments are necessary while models remain sample-inefficient, but deployment data could matter much more as learning improves.
- Off-policy learning offers a way to learn from real interaction traces without replaying every interaction in simulation.